跳转至

分久必合：三大 LLM Agent 框架的架构趋同

文章背景与核心概要

“Agent Harness”（代理框架）是指将基础大语言模型转化为自主智能体的周边软件基础设施。随着技术演进，这一运行时层已逐渐取代模型本身，成为制约智能体能力与行为表现的主要瓶颈。本文通过对比 LangChain 的 deepagents、Earendil 的 pi 以及 DeepSeek 的 dsh 这三个设计哲学迥异的开源框架，深入探讨了 Agent 架构的演变规律。

研究发现，尽管这些框架的起源路径各异，但它们在演进过程中独立地收敛于一套包含五大核心组件的统一架构标准。此外，文章还揭示了当前所有主流框架在“外部可验证性”方面的共同缺失，并指出这正是未来高信任度领域 Agent 框架发展的关键前沿方向。

执行摘要

“Agent harness” 指的是将基础语言模型转化为自主智能体的周边软件基础设施。这涉及构建上下文、管理工具、执行循环以及在长时间运行中维护状态。目前，这一运行时层（而非底层语言模型本身）正日益成为智能体能力和行为的主要瓶颈。

本研究考察了三个基于截然不同设计哲学构建的开源编码智能体框架： 1. LangChain 的 deepagents（功能完备型） 2. Earendil 的 pi（极简主义型） 3. DeepSeek 的 dsh（一切皆插件型）

尽管这些框架源自完全相反的设计轨迹——成熟框架要么在舍弃不必要的脚手架，要么在逐步构建持久的基础设施——但它们已独立收敛于一个由五个核心组件组成的统一架构标准。此外，本文还指出了所有受检系统中的一个显著预测性差距：外部可验证性。

An agent harness refers to the surrounding software infrastructure that transforms a foundational language model into an autonomous agent. This involves constructing context, managing tools, executing loops, and maintaining state over extended execution runs. Increasingly, this runtime layer—rather than the underlying language model itself—acts as the primary bottleneck for agent capability and behavior.

This study examines three open-source coding-agent harnesses built from contrasting design philosophies: 1. LangChain's deepagents (batteries-included) 2. Earendil's pi (radical minimalism) 3. DeepSeek's dsh (everything-is-a-plugin)

Despite originating from opposite design trajectories—with mature frameworks either discarding unnecessary scaffolding or progressively building durable infrastructure—the harnesses have independently converged on a unified architectural standard consisting of five core components. Furthermore, the paper identifies a notable predictive gap across all examined systems: external verifiability.

关键发现与架构趋同

通过源代码级分析和历史提交记录追踪，本研究证明了多样化的智能体框架正自然地演变为一种包含五个重复元素的共享中间形态架构：

  1. 商品化循环（Commoditised Loop）： 编排模型调用和工具响应的标准化执行流程。
  2. 仅追加的可重放会话记录（Append-Only Replayable Session Record）： 确保状态持久性和执行可追溯性的不可变、按时间顺序排列的日志。
  3. 模型特性数据化（Model Quirks Kept as Data）： 通过将模型特定的怪癖和变通逻辑视为外部化数据，将其与核心执行代码解耦。
  4. 上下文的渐进式披露（Progressive Disclosure of Context）： 动态上下文管理，通过增量揭示信息以防止 Token 窗口饱和。
  5. 显式扩展接口（Explicit Extension Seams）： 定义明确的模块化接口，允许插件、工具和自定义行为进行简洁集成。

研究对第三个独立框架（dsh）进行了评估以测试该模型，结果显示其具备所有五个趋同标志，甚至直接复用了其他框架的实现组件。作者将这种架构统一归因于平行发现、生态扩散和字面代码复用的结合。

Through source-level analysis and historical commit tracking, the study demonstrates that diverse agent frameworks naturally evolve toward a shared middle-form architecture comprising five recurring elements:

  1. A Commoditised Loop: Standardized execution flow orchestrating model calls and tool responses.
  2. An Append-Only Replayable Session Record: Immutable, chronological logs ensuring state persistence and execution traceability.
  3. Model Quirks Kept as Data: Decoupling model-specific idiosyncrasies and workaround logic from core execution code by treating them as externalized data.
  4. Progressive Disclosure of Context: Dynamic context management that reveals information incrementally to prevent token window saturation.
  5. Explicit Extension Seams: Well-defined modular interfaces allowing plugins, tools, and custom behaviors to integrate cleanly.

A third held-out framework (dsh) was evaluated to test this model and exhibited all five convergence markers, even literally reusing an implementation component from another harness. The author attributes this architectural unification to a combination of parallel discovery, ecosystem diffusion, and literal code reuse.

盲点：外部可验证性

虽然内部一致性和可重放性得到了很好的支持，但所有被调查的框架中仍完全缺失一个关键维度：外部可验证性。

本文将其定义为一种防篡改记录，即第三方可以在不信任本地运行时的情况下，通过密码学或逻辑方式进行审计。研究并未将这一缺失视为疏忽，而是将其定义为一个预测性差距——代表了未来为来源敏感和高信任领域设计的智能体框架将产生分歧的前沿轴线。

While internal consistency and replayability are well-supported, one critical dimension remains completely absent across all investigated harnesses: external verifiability.

The paper defines this as a tamper-evident record that an independent third party can cryptographically or logically audit without placing trust in the local runtime. Rather than viewing this omission as an oversight, the study frames it as a predictive gap—representing the frontier axis where future agent harnesses designed for provenance-sensitive and high-trust domains will diverge.

访问与资源